1. science
2. mathematics
3. algebra

CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
Abstract. The term “closed form” is one of those mathematical notions that is
commonplace, yet virtually devoid of rigor. And, there is disagreement even on
the intuitive side; for example, most everyone would say that π + log 2 is a closed
form, but some of us would think that the Euler constant γ is not closed. Like
others before us, we shall try to supply some missing rigor to the notion of closed
forms and also to give examples from modern research where the question of closure
looms both important and elusive.
1. Closed Forms: What They Are
Mathematics abounds in terms which are in frequent use yet which are rarely
made precise. Two such are rigorous proof and closed form (absent the technical
use within differential algebra). If a rigorous proof is “that which ‘convinces’ the
appropriate audience” then a closed form is “that which looks ‘fundamental’ to the
requisite consumer.” In both cases, this is a community-varying and epoch-dependent
notion. What was a compelling proof in 1810 may well not be now; what is a fine
closed form in 2010 may have been anathema a century ago. In the article we are
intentionally informal as befits a topic that intrinsically has no one “right” answer.
Let us begin by sampling the Web for various approaches to informal definitions
of “closed form.”
1.0.1. First approach to a definition of closed form. The first comes from
MathWorld [55] and so may well be the first and last definition a student or other
seeker-after-easy-truth finds.
An equation is said to be a closed-form solution if it solves a given problem
in terms of functions and mathematical operations from a given generally
accepted set. For example, an infinite sum would generally not be considered
closed-form. However, the choice of what to call closed-form and what not is
Date: October 22, 2010.
Borwein:
Centre for Computer Assisted Research Mathematics and its Applications (CARMA), University of Newcastle, Callaghan, NSW 2308, Australia.
Email:
jonathan.borwein@newcastle.edu.au.
Crandall: Center for Advanced Computation, Reed College, Portland OR, USA. Email:
crandall@reed.edu.
1
2
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
rather arbitrary since a new “closed-form” function could simply be defined
in terms of the infinite sum.—Eric Weisstein
There is not much to disagree with in this but it is far from rigorous.
1.0.2. Second approach. The next attempt follows a 16 September 1997 question
to the long operating “Dr. Math.” site1 and is a good model of what interested
students are likely to be told.
Subject: Closed form solutions
Dear Dr. Math, What is the exact mathematical definition of a closed
form solution? Is a solution in “closed form” simply if an expression relating
all of the variables can be derived for a problem solution, as opposed to some
higher-level problems where there is either no solution, or the problem can
only be solved incrementally or numerically?
Sincerely, . . . .
The answer followed on 22 Sept:
This is a very good question! This matter has been debated by mathematicians for some time, but without a good resolution.
Some formulas are agreed by all to be “in closed form.” Those are the ones
which contain only a finite number of symbols, and include only the operators +, −, ∗, /, and a small list of commonly occurring functions such as n-th
roots, exponentials, logarithms, trigonometric functions, inverse trigonometric functions, greatest integer functions, factorials, and the like.
More controversial would be formulas that include infinite summations
or products, or more exotic functions, such as the Riemann zeta function,
functions expressed as integrals of other functions that cannot be performed
symbolically, functions that are solutions of differential equations (such as
Bessel functions or hypergeometric functions), or some functions defined
recursively. Some functions whose values are impossible to compute at some
specific points would probably be agreed not to be in closed form (example:
f (x) = 0 if x is an algebraic number, but f (x) = 1 if x is transcendental.
For most numbers, we do not know if they are transcendental or not). I
hope this is what you wanted.
No more formal, but representative of many dictionary definitions is:
1.0.3. Third approach. A coauthor of the current article is at least in part responsible for the following brief definition from a recent mathematics dictionary [16]:
closed form n. an expression for a given function or quantity, especially
an integral, in terms of known and well understood quantities, such as the
1
Available at http://mathforum.org/dr/math/.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
evaluation of
as
√
Z
3
∞
exp(−x2 ) dx
−∞
π.—Collins Dictionary
And of course one cares more for a closed form when the object under study is
important, such as when it engages the normal distribution as above.
With that selection recorded, let us turn to some more formal proposals.
1.0.4. Fourth approach. Various notions of elementary numbers have been proposed.
Definition [30]. A subfield F of C is closed under exp and log if (1)
exp(x) ∈ F for all x ∈ F and (2) log(x) ∈ F for all nonzero x ∈ F ,
where log is the branch of the natural logarithm function such that −π <
Im(log x) ≤ π for all x. The field E of EL numbers is the intersection of all
subfields of C that are closed under exp and log.—Tim Chow
Tim Chow explains nicely why he eschews capturing all algebraic numbers in his
definition; why he wishes only elementary quantities to have closed forms; whence he
prefers E to Ritt’s 1948 definition of elementary numbers as the smallest algebraically
closed subfield L of C that is closed under exp and log. His reasons include that:
Intuitively, “closed-form” implies “explicit,” and most algebraic functions have no simple explicit expression.
Assuming Shanuel’s conjecture that given n complex numbers z1 , ..., zn which are
linearly independent over the rational numbers, the extension field
Q (z1 , ..., zn , ez1 , ..., ezn )
has transcendence degree of at least n over the rationals, then the algebraic members
of E are exactly those solvable in radicals [30]. We may thence think of Chow’s class
as the smallest plausible class of closed forms. Only a mad version of Markov would
want to further circumscribe the class.
1.1. Special functions. In an increasingly computational world, an explicit/implicit
dichotomy is occasionally useful; but not very frequently. Often we will prefer computationally the numerical implicit value of an algebraic number to its explicit tower
of radicals; and it seems increasingly perverse to distinguish the root of 2x5 − 10x + 5
from that of 2x4 − 10x + 5 or to prefer arctan(π/7) to arctan(1). We illustrate these
issues further in Example 3.1, 3.3 and 4.3.
We would prefer to view all values of classical special functions of mathematical
physics [53] at algebraic arguments as being closed forms. Again there is no generally accepted class of special functions, but most practitioners would agree that
the solutions to the classical second-order algebraic differential equations (linear or
4
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
say Painlev´e) are included. But so are various hypertranscendental functions such as
Γ, B and ζ which do not arise in that way.2
Hence, we do not wish to accept any definition of special function which relies on
the underlying functions satisfying natural differential equations. The class must be
extensible, new special functions are always being discovered.
A recent American Mathematical Monthly review3 of [46] says:
There’s no rigorous definition of special functions, but the following definition is in line with the general consensus: functions that are commonly used
in applications, have many nice properties, and are not typically available
on a calculator. Obviously this is a sloppy definition, and yet it works fairly
well in practice. Most people would agree, for example, that the Gamma
function is included in the list of special functions and that trigonometric
functions are not.
Once again, there is much to agree with, and much to quibble about, in this reprise.
That said, most serious books on the topic are little more specific. For us, special
functions are non-elementary functions about which a significant literature has developed because of their importance in either mathematical theory or in practice.
We certainly prefer that this literature includes the existence of excellent algorithms
for their computation. This is all consonant with—if somewhat more ecumenical
than—Temme’s description in the preface of his lovely book [53, Preface p. xi]:
[W]e call a function “special” when the function, just like the logarithm,
the exponential and trigonometric functions (the elementary transcendental functions), belongs to the toolbox of the applied mathematician, the
physicist or engineer.
Even more economically, Andrews, Askey and Roy start the preface to their important book Special functions [1] by writing:
Paul Turan once remarked that special functions would be more appropriately labeled “useful functions.”
With little more ado, they then start to work on the Gamma and Beta functions;
indeed the term “special function” is not in their index. Near the end of their preface,
they also write
[W]e suggest that the day of formulas may be experiencing a new dawn.
Example 1.1 (Lambert’s W). The Lambert W function, W (x), is defined by appropriate solution of y · ey = x [20, pp. 277–279]. This function has been implemented
in computer algebra systems (CAS); and has many uses despite being unknown to
most scientists and only relatively recently named [40]. It is now embedded as a
2Of
course a value of an hypertranscendental function at algebraic argument may be very well
behaved, see Example 1.4.
3Available at http://mathdl.maa.org/mathDL/19/?pa=reviews&sa=viewBook&bookId=69684.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
(a) modulus of W
5
(b) W on real line
Figure 1. The Lambert W function.
primitive in Maple and Mathematica with the same status as any other well studied
special or elementary function. (See for example the tome [25].) The CAS know its
power series and much more. For instance in Maple entering:
> fsolve(exp(x)*x=1);identify(%);
returns
0.5671432904, LambertW(1)
We consider this to be a splendid closed form even though, again assuming Shanuel’s
conjecture, W (1) 6∈ E [30]. Additionally, it is only recently rigorously proven that
W is not an elementary function in Liouville’s precise sense [25]. We also note that
successful simplification in a modern CAS [28] requires a great deal of knowledge of
special functions.
1.2. Further approaches.
1.2.1. Fifth approach. PlanetMath’s offering, as of 15 February 20104, is certainly
in the elementary number corner.
expressible in closed form (Definition) An expression is expressible in
a closed form, if it can be converted (simplified) into an expression containing only elementary functions, combined by a finite amount of rational
operations and compositions.—Planet Math
4Available
at http://planetmath.org/encyclopedia/ClosedForm4.html.
6
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
This reflects both much of what is best and what is worst about ‘the mathematical
wisdom of crowds’. For the reasons adduced above, we wish to distinguish—but
admit both—those closed forms which give analytic insight from those which are
sufficient and prerequisite for effective computation. Our own current preferred class
[7] is described next.
1.2.2. Sixth approach. We wish to establish a set X of generalized hypergeometric evaluations; see [7] for an initial, rudimentary definition which we shall refine
presently. First, we want any convergent sum
X
x=
cn z n
(1.1)
n≥0
to be an element of our set X, where z is algebraic, c0 is rational, and for n > 0,
A(n)
cn−1
cn =
B(n)
for integer polynomials A, B with deg A ≤ deg B. Under these conditions the expansion for x converges absolutely on the open disk |z| < 1. However, we also allow
x to be any finite analytic-continuation value of such a series; moreover, when z lies
on a branch cut we presume both branch limits to be elements of X. (See ensuing
examples for some clarification.) It is important to note that our set X is closed
under rational multiplication, due to freedom of choice for c0 .
Example 1.2 (First members of X). The generalized hypergeometric function evaluation
a1 , . . . ap+1 z
p+1 Fp
b1 , . . . , b p for rational ai , bj with all bj positive has branch cut z ∈ (0, ∞), and the evaluation
is an element of X for complex z not on the cut (and the evaluation on each side of
said cut is also in X).
P
The trilogarithm Li3 (z) := n≥1 z n /n3 offers a canonical instance. Formally,
1
1, 1, 1, 1 Li3 (z) = 4 F3
z .
z
2, 2, 2 and for z = 1/2 the hypergeometric series converges absolutely, with
1
7
1
π2
Li3
= ζ(3) + log3 2 −
log 2.
2
8
6
12
Continuation values at z = 2 on the branch cut can be inferred as
7
π2
π
log 2 ± i log 2,
lim+ Li3 (2 ± i) = ζ(3) +
→0
16
8
4
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
7
so both complex
numbers on the right here are elements of X. The quadralogarithmic
value Li4 21 is thought not to be similarly decomposable but likewise belongs to
X.
Now we are prepared to posit
Definition [7]. The ring of hyperclosure H is the smallest subring of C
containing the set X. Elements of H are deemed hyperclosed.
In other words, the ring H is generated by all general hypergeometric evaluations,
under the ·, + operators, all symbolized by
H = hXi·,+ .
H will contain a great many interesting closed forms from modern research. Note
that H contains all closed forms in the sense of Wilf and Zeilberger [47, Ch. 8]
wherein only finite linear combinations of hypergeometric evaluations are allowed.
So what numbers are in the ring H? First off, almost no complex numbers belong
to this ring! This is easily seen by noting that the set of general hypergeometric
evaluations is countable, so the generated ring must also be countable. Still, a great
many fundamental numbers are provably hyperclosed. Examples follow, in which we
let ω denote an arbitrary algebraic number and n any positive integer:
ω, log ω, eω , π
1
the dilogarithmic combination Li2 √
+ (log 2)(log 3),
5
the elliptic integral K(ω),
the zeta function values ζ(n),
special functions such as the Bessel evaluations Jn (ω).
Incidentally, it occurs in some modern experimental developments that the real
or imaginary part of a hypergeometric evaluation is under scrutiny. Generally, <, =
operations preserve hyperclosure, simply because a the series (or continuations) at
z and z ∗ can be linearly combined in the ring H. Referring to Example 1.2, we
see that for algebraic z, the number < (Li3 (z)) is hyperclosed;and even
on
the
2
7
z
ζ(3) + π8 log 2 is hyperclosed. In general, < p+1 Fp ···
is
cut, < (Li3 (2)) = 16
··· hyperclosed.
We are not claiming that hyperclosure is any kind of final definition for “closed
forms.” But we do believe that any defining paradigm for closed forms must include this ring of hyperclosure H. One way to reach further is to define a ring of
superclosure as the closure
S := hHH i·,+ .
8
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
This ring contains numbers such as
eπ + π e ,
1
,
ζ(3)ζ(5)
and of course a vast collection of numbers that may not belong to H itself. If we say
that an element of S is superclosed, we still preserve the countability of all superclosed
numbers. Again, any good definition of “closed form” should incorporate whatever
is in the ring S.
1.3. Seventh approach. In a more algebraic topological setting, it might make
sense to define closed forms to be those arising as periods—that is as integrals of
rational functions (with integer parameters) in n variables over domains defined
by algebraic equations. These ideas originate in the theory of elliptic and abelian
integrals and are deeply studied [41]. Periods form an algebra and certainly capture
many constants. They are especially well suited to the study of L-series, multi zeta
values, polylogarithms and the like, but again will not capture all that we wish. For
example, e is conjectured not to be a period, as is Euler’s constant γ (see Section 5).
Moreover, while many periods have nice series, it is not clear that all do.
As this takes us well outside our domain of expertise we content ourselves with two
examples originating in the study of Mahler measures. We refer to a fundamental
paper by Denninger [38] and a very recent paper of Rogers [49] for details.
Example 1.3 (Periods and Mahler Measures [38]). The logarithmic Mahler measure
of a polynomial P in n-variables can be defined as
Z 1
Z 1Z 1
log P e2πiθ1 , · · · , e2πiθn dθ1 · · · dθn .
···
µ(P ) :=
0
0
0
Then µ(P ) turns out to be an example of a period and its exponential, M (P ) :=
exp(µ(P )), is a mean of the values of P on the unit n-torus. When n = 1 and
P has integer coefficients M (P ) is always an algebraic integer. An excellent online
synopsis can be found in Dave Boyd’s article http://eom.springer.de/m/m120070.
htm. Indeed, Boyd has been one of the driving forces in the field. A brief introduction
to the univariate case is also given in [21, 358–359].
There is a remarkable series of recent results—many more discovered experimentally than proven—expressing various multi-dimensional µ(P ) as arithmetic quantities. Boyd observes that there appears to be a tight connection to K-theory. An
0
early result due to Smyth (see [50], also [52]) is that µ(1 + x + y) = L3 (−1). Here
L3 is the Dirichlet L-series modulo three. A partner result of Smyth’s is is that
µ(1 + x + y + z) = 7 ζ(3)/π 2 , a number that is certainly hyperclosed, being as both
ζ(3) and 1/π are. A conjecture of Denninger [38], confirmed to over 50 places, is
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
9
that
15
LE (2)
(1.2)
π2
is an L-series value over an elliptic curve E with conductor 15. Rogers [49] recasts
(1.2) as
2
∞ 15 X 2n (1/16)2n+1
?
F (3, 5) =
,
(1.3)
π 2 n=0 n
2n + 1
?
µ(1 + x + y + 1/x + 1/y) =
where
F (b, c) := (1 + b)(1 + c)
(−1)n+m+j+k
((6n + 1)2 + b(6m + 1)2 + c(6j + 1)2 + bc(6k + 1)2 )2
n,m,j,k
X
is a four-dimensional lattice sum.
While (1.3) remains a conjecture, Rogers is able to evaluate many values of F (b, c)
in terms of Meijer-G or hypergeometric functions. We shall consider the most famous
crystal sum, the Madelung constant, in Example 5.1.
It is striking how beautiful combinatorial games can be when played under the
rubric of hyper- or superclosure.
Example 1.4 (Superclosure of rational values of Γ). Let us begin with the Beta function
Γ(r)Γ(s)
B(r, s) :=
.
Γ(r + s)
R∞
with Γ(s) defined if one wishes as Γ(s) := 0 ts−1 e−t dt. It turns out that for
any rationals r, s the Beta function is hyperclosed. This is immediate from the
hypergeometric identities
1
rs
−r, −s =
1
2 F1
B(r, s)
r+s
1 π sin π(r + s) (1 − r)M (1 − s)M
r, s B(r, s) =
1 ,
2 F1
sin πr sin πs M !(1 − r − s)M
M + 1
where M is any integer chosen such that the hypergeometric series converges, say
M = d1 + r + se. (Each of these Beta relations is a variant of the celebrated Gauss
evaluation of 2 F1 at 1 [1, 53] and is also the reason B is a period.)
We did not seize upon the Beta function arbitrarily, for, remarkably, the hyperclosure of B ±1 (r, s) leads to compelling results on the Gamma function itself. Indeed,
consider for example this product of four Beta-function evaluations:
Γ(1/5)Γ(1/5) Γ(2/5)Γ(1/5) Γ(3/5)Γ(1/5) Γ(4/5)Γ(1/5)
·
·
·
.
Γ(2/5)
Γ(3/5)
Γ(4/5)
Γ(5/5)
10
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
We know this product is hyperclosed. But upon inspection we see that the product
is just Γ5 (1/5). Along such lines one can prove that for any positive rational a/b (in
lowest terms), we have hyperclosure of powers of the Gamma-function, in the form:
Γ±b (a/b) ∈ H.
Perforce, we have therefore a superclosure result for any Γ(rational) and its reciprocal:
Γ±1 (a/b) ∈ S.
b
Again like calculations
1 show Γ (a/b) is a period [41]. One fundamental consequence
−2 1
= π is hyperclosed; thus every integer power of π is hyperclosed.
is thus: Γ
2
Incidentally, deeper combinatorial analysis shows that—in spite of our Γ5 51 Betachain above, it really only takes logarithmically many (i.e., O(log b)) hypergeometric
evaluations to write Gamma-powers. For example,
1 1 4
2 2 2
4 4 1
1
− 7 , − 7 − 7 , − 7 − 7 , − 7 −7
Γ
= 3 6 2 F1
1
F
1
F
1 .
2
1
2
1
7
27
1 1 1 We note also that for Γ(n/24) with n integer, elliptic integral algorithms are known
which converge as fast as those for π [26, 21].
The above remarks on superclosure of Γ(a/b) lead to the property of superclosure
for special functions such as Jν (ω) for algebraic ω and rational ν; and for many of
the mighty Meijer-G functions, as the latter can frequently be written by Slater’s
theorem [14] as superpositions of hypergeometric evaluations with composite-gamma
products as coefficients. (See Example 3.2 below for instances of Meijer-G in current
research.)
There is an interesting alternative way to envision hyperclosure, or at least something very close to our above definition. This is an idea of J. Carette [27], to the effect that solutions at algebraic end-points, and algebraic initial points, for holonomic
ODEs—i.e. differential-equation systems with integer-polynomial coefficients—could
be considered closed. One might say diffeoclosed. An example of a diffeoclosed number is J1 (1), i.e., from the Bessel differential equation for J1 (z) with z ∈ [0, 1]; it
suffices without loss of generality to consider topologically clean trajectories of the
variable over [0, 1]. There is a formal ring of diffeoclosure, which ring is very similar
to our H; however there is the caution that trajectory solutions can sometimes have
nontrivial topology, so precise ring definitions would need to be effected carefully.
It is natural to ask “what is the complexity of hypergeometric evaluations?” Certainly for the converging forms with variable z on the open unit disk, convergence is
geometric, requiring O(D1+ ) operations to achieve D good digits. However, in very
many cases this can be genuinely enhanced to O(D1/2+ ) [21].
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
11
2. Closed Forms: Why We Care
In many optimization problems, simple, approximate solutions are more
useful than complex exact solutions.—Steve Wright
As Steve Wright observed in a recent lecture on sparse optimization it may well
be that a complicated analytic solution is practically intractable but a simplifying
assumption leads to a very practical closed form approximation (e.g., in compressed
sensing). In addition to appealing to Occam’s razor, Wright instances that:
(a)
(b)
(c)
(d)
the data quality may not justify exactness;
the simple solution may be more robust;
it may be easier to “explain/ actuate/ implement/ store”;
and it may conform better to prior knowledge.
As mathematical discovery more and more involves extensive computation, the
premium on having a closed form increases. The insight provided by discovering a
closed form ideally comes at the top of the list but efficiency of computation will run
a good second.
Example 2.1 (The amplitude of a pendulum). Wikipedia 5 after giving the classical
small angle (simple harmonic) approximation
s
L
p ≈ 2π
g
for the period p of a pendulum of length L and amplitude α, develops the exact
solution in a form equivalent to
s
L α
p=4
K sin
g
2
and then says:
This integral cannot be evaluated in terms of elementary functions. It can
be rewritten in the form of the elliptic function of the first kind (also see
Jacobi’s elliptic functions), which gives little advantage since that form is
also insoluble.
True, an elliptic-integral solution is not elementary, yet the notion of insolubility is
misleading for two reasons: First, it is known that for some special angles α, the
pendulum period can be given a closed form. As discussed in [32], one exact solution
5Available
at http://en.wikipedia.org/wiki/Pendulum_(mathematics).
12
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
is, for α = π/2 (so pendulum is released from horizontal-rod position),
s ! √
L
π
.
p = 2π
2
g Γ (3/4)
It
measurable in even a rudimentary laboratory that the excess factor here,
√ is readily
πΓ−2 (3/4) ≈ 1.18034 looks just right, i.e., a horizontal-release pendulum takes 18
per cent longer to fall. Moreover, there is an exact dynamical solution for the timedependent angle α(t); namely for a pendulum with α(±∞) = ±π and α(0) = 0, i.e.
the bob crosses angle zero (hanging straight down) at time zero, but in the limits of
time → ±∞ the bob ends up straight vertical. We have period p = ∞, yet the exact
angle α(t) for given t can be written down in terms of elementary functions!
The second misleading aspect is this: K is—for any α—remarkably tractable in a
computational sense. Indeed K admits of a quadratic transformation
√
1 − 1 − k2
√
K (k) = (1 + k1 ) K (k1 ) , k1 :=
(2.1)
1 + 1 − k2
as was known already to Landen, Legendre and Gauss.
In fact all elementary function to very high precision are well computed via K [21].
So the comment was roughly accurate in the world of slide rules or pocket calculators; it is misleading today—if one has access to any computer package. Nevertheless,
both deserve to be called closed forms: one exact and the other an elegant approximate closed form (excellent in its domain of applicability, much as with Newtonian
mechanics) which is equivalent to
α π
K sin
≈
2
2
for small initial amplitude α. To compute K(π/6) = 1.699075885 . . . to five places
requires using (2.1) only twice and then estimating the resultant integral by π/2. A
third step gives the ten-digit precision shown.
It is now the case that much mathematical computation is hybrid : mixing numeric
and symbolic computation. Indeed, which is which may not be clear to the user if,
say, numeric techniques have been used to return a symbolic answer or if a symbolic
closed form has been used to make possible a numerical integration. Moving from
classical to modern physics, both understanding and effectiveness frequently demand
hybrid computation.
Example 2.2 (Scattering amplitudes [2]). An international team of physicists, in
preparation for the Large Hadron Collider (LHC), is computing scattering amplitudes
involving quarks, gluons and gauge vector bosons, in order to predict what results
could be expected on the LHC. By default, these computations are performed using
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
13
conventional double precision (64-bit IEEE) arithmetic. Then if a particular phase
space point is deemed numerically unstable, it is recomputed with double-double
precision. These researchers expect that further optimization of the procedure for
identifying unstable points may be required to arrive at an optimal compromise
between numerical accuracy and speed of the code. Thus they plan to incorporate
arbitrary precision arithmetic, into these calculations. Their objective is to design
a procedure where instead of using fixed double or quadruple precision for unstable
points, the number of digits in the higher precision calculation is dynamically set
according to the instability of the point. Any subroutine which uses a closed form
symbolic solution (exact or approximate) is likely to prove much more robust and
efficient.
3. Detailed Examples
We start with three examples originating in [15].
In the January 2002 issue of SIAM News, Nick Trefethen presented ten diverse
problems used in teaching modern graduate numerical analysis students at Oxford
University, the answer to each being a certain real number. Readers were challenged
to compute ten digits of each answer, with a $100 prize to the best entrant. Trefethen
wrote,
“If anyone gets 50 digits in total, I will be impressed.”
To his surprise, a total of 94 teams, representing 25 different nations, submitted
results. Twenty of these teams received a full 100 points (10 correct digits for each
problem). The problems and solutions are dissected most entertainingly in [15]. One
of the current authors wrote the following in a review [18] of [15].
Success in solving these problems required a broad knowledge of mathematics and numerical analysis, together with significant computational effort, to
obtain solutions and ensure correctness of the results. As described in [15]
the strengths and limitations of Maple, Mathematica, Matlab (The 3Ms),
and other software tools such as PARI or GAP, were strikingly revealed in
these ventures. Almost all of the solvers relied in large part on one or more
of these three packages, and while most solvers attempted to confirm their
results, there was no explicit requirement for proofs to be provided.
Example 3.1 (Trefethen problem #2 [15, 18]).
A photon moving at speed 1 in the x-y plane starts at t = 0 at (x, y) =
(1/2, 1/10) heading due east. Around every integer lattice point (i, j)
in the plane, a circular mirror of radius 1/3 has been erected. How far
from the origin is the photon at t = 10?
14
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
Using interval arithmetic with starting intervals of size smaller than 10−5000 , one
can actually find the position of the particle at time 2000—not just time ten. This
makes a fine exercise in very high-precision interval computation, but in absence
of any closed form one is driven to such numerical gymnastics to deal with error
propagation.
Example 3.2 (Trefethen’s problem #9 [15, 18]).
R2
The integral I(a) = 0 [2 + sin(10α)]xα sin(α/(2 − x)) dx depends on
the parameter α. What is the value α ∈ [0, 5] at which I(α) achieves
its maximum?
The maximum parameter is expressible in terms of a Meijer-G function which is
a special function with a solid history. While knowledge of this function was not
common among the contestants, Mathematica and Maple both will figure this out
[14], and then the help files or a web search will quickly inform the scientist.
This is another measure of the changing environment. It is usually a good idea—
and not at all immoral—to data-mine. These Meijer-G functions, first introduced in
1936, also occur in quantum field theory and many other places [8]. For example,
the moments of an n-step random walk in the plane are given for s > 0 by
n
s
Z
X
2πxk i Wn (s) :=
e
(3.1)
dx.
[0,1]n k=1
It transpires [23, 35] that for all complex s
Γ(1 + s/2)
2,1
G3,3
W3 (s) =
Γ(1/2)Γ(−s/2)
1, 1, 1 1
,
1
, − 2s , − 2s 4
2
(3.2)
Moreover, for s not an odd integer, we have
1 1 1 s s s πs s 2
, , − 2 , − 2 , − 2 1
s
1
2 2 2 1
W3 (s) = 2s+1 tan
+
F
3 F2
3
2
s−1
s
s+3 s+3 4 .
2
2
, 2 4
1, − s−1
2
2
2
2
We have not given the somewhat technical definition of MeijerG, but Maple, Mathematica, Google searches, Wikipedia, the DLMF or many other tools will.
There are two corresponding formulae for W4 . We thus know, from our “Sixth
approach” section previous in regard to superclosure of Γ-evaluations, that both
W3 (q), W4 (q) are superclosed for rational argument q for q not an odd integer. We
illustrate by showing graphs of W3 , W4 on the real line in Figure 2 and in the complex
plane in Figure 3. The later highlights the utility of the Meijer-G representations.
Note the poles and removable singularities.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
!6
!4
15
4
4
3
3
2
2
1
1
!2
2
!6
!4
!2
2
!1
!1
!2
!2
!3
!3
(a) W3
(b) W4
Figure 2. Moments of n-step walks in the plane. W3 , W4 analytically
continued to the real line.
(a) W3
(b) W4
Figure 3. W3 via (3.2) and W4 in the complex plane.
The Meijer-G functions are now described in the newly completed Digital Library
of Mathematical Functions 6 and as such are now full, indeed central, members of the
family of special functions.
Example 3.3 (Trefethen’s problem #10 [15, 18]).
A particle at the center of a 10 × 1 rectangle undergoes Brownian
motion (i.e., 2-D random walk with infinitesimal step lengths) till it
6A
massive revision of Abramowitz and Stegun—with the now redundant tables removed, it is
available at www.dlmf.nist.gov. The hard copy version is also now out [44]. It is not entirely a
substitute for the original version as coverage has changed.
16
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
hits the boundary. What is the probability that it hits at one of the
ends rather than at one of the sides?
Hitting the Ends. Bornemann [15] starts his remarkable solution by exploring
Monte-Carlo methods, which are shown to be impracticable. He then reformulates
the problem deterministically as the value at the center of a 10 × 1 rectangle of an
appropriate harmonic measure [56] of the ends, arising from a 5-point discretization
of Laplace’s equation with Dirichlet boundary conditions. This is then solved by
a well chosen sparse Cholesky solver. At this point a reliable numerical value of
3.837587979 · 10−7 is obtained. And the posed problem is solved numerically to the
requisite ten places.
This is the warm up. We may proceed to develop two analytic solutions, the first
using separation of variables on the underlying PDE on a general 2a × 2b rectangle.
We learn that
∞
4 X (−1)n
π(2n + 1)
p(a, b) =
sech
ρ
(3.3)
π n=0 2n + 1
2
where ρ := a/b. A second method using conformal mappings, yields
arccot ρ = p(a, b)
π
+ arg K eip(a,b)π
2
(3.4)
where K is again the complete elliptic integral of the first kind. It will not be apparent
to a reader unfamiliar with inversion of elliptic integrals that (3.3) and (3.4) encode
the same solution—though they must as the solution is unique in (0, 1)—and each
can now be used to solve for ρ = 10 to arbitrary precision. Bornemann ultimately
shows that the answer is
p=
2
arcsin (k100 ) ,
π
(3.5)
where
k100
2 2
√
√ √ √ √
4
3 − 2 2 2 + 5 −3 + 10 − 2 + 5
:=
.
No one (except harmonic analysts perhaps) anticipated a closed form—let alone one
like this.
Where does this come from? In fact [21, (3.2.29)] shows that
∞
X
(−1)n
π(2n + 1)
1
sech
ρ = arcsin k,
2n + 1
2
2
n=0
(3.6)
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
17
exactly when kρ2 is parameterized by theta functions in terms of the so called nome,
q = exp(−πρ), as Jacobi discovered. We have
P∞
(n+1/2)2
θ22 (q)
n=−∞ q
P∞
=
,
q := e−πρ .
(3.7)
kρ2 = 2
2
n
θ3 (q)
n=−∞ q
Comparing (3.6) and (3.3) we see that the solution is
k100 = 6.02806910155971082882540712292 . . . · 10−7
as asserted in (3.5).
The explicit form now follows from classical nineteenth century theory as discussed
say in [15, 21]. In fact k210 is the singular value sent by Ramanujan to
√ Hardy in his
famous letter of introduction
[20, 21]. If Trefethen had asked for a 210 × 1 box,
√
√
or even better a 15 × 14 one, this would have shown up in the answer since in
general
2
p(a, b) = arcsin ka2 /b2 .
(3.8)
π
Alternatively, armed only with the knowledge that the singular values of rational
parameters are always algebraic we may finish entirely computationally as described
in [18].
(a) A Superior Mirage
(b) An Inferior Mirage
Figure 4. Two Impressive Mirages
We finish this section with two attractive applied examples from optics and astrophysics respectively.
Example 3.4 (Mirages [45]). In [45] the authors, using geometric methods, develop an
exact but implicit formula for the path followed by a light ray propagating over the
earth with radial variations in the refractive index. By suitably simplifying they are
18
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
able to provide an explicit integral closed form. They then expand it asymptotically.
This is done with the knowledge that the approximation is good to six or seven
places—more than enough to use it on optically realistic scales. Moreover, in the
case of quadratic or linear refractive indices these steps may be done analytically.
In other words, as advanced by Wright, a tractable and elegant approximate closed
form is obtained to replace a problematic exact solution. From these forms interesting qualitative consequences follow. With a quadratic index, images are uniformly
magnified in the vertical direction; only with higher order indices can nonuniform
vertical distortion occur. This sort of knowledge allows one, for example, to correct
distortions of photographic images as in Figure 4, with confidence and efficiently. Example 3.5 (Structure of stars). The celebrated Lane–Emden equation, presumed
to describe the pressure χ at radius r within a star, can be put in the form:
d2 χ
= −χn ,
(3.9)
2
dr
with boundary conditions χ(0) = 0, χ0 (0) = 1, and positive real constant n, all of
this giving rise to a unique trajectory χn (r) on r ∈ [0, ∞). (Some authors invoke
the substitution χ(r) := rθ(r) to get an equivalent ODE for temperature θ; see [29]).
The beautiful thing is, where this pressure trajectory crosses zero for positive radius
r is supposed to be the star radius; call that zero zn .
Amazingly, the Lane–Emden equation has known exact solutions for n = 0, 1, 5.
The pressure trajectories for which indices n being respectively
1
χ0 (r) = − r3 + r,
(3.10)
6
χ1 (r) = sin r,
(3.11)
r
χ5 (r) = p
.
(3.12)
1 + r2 /3
√
The respective star radii are thus closed forms z0 = 6 and z1 = π, while for (3.12)
with index n = 5 we have infinite star radius (no positive zero for the pressure χ5 ).
In the spirit of our previous optics example, the Lane–Emden equation is a simplification of a complicated underlying theory—in this astrophysics case, hydrodynamics—
and one is rewarded by some closed-form star radii. But what about, say, index
n = 2? We do not know a closed-form function for the χ trajectory in any convenient sense. What the present authors have calculated (in 2005) is the n = 2 star
radius, as a high-precision number
rn−1
z2 = 4.352874595946124676973570061526142628112365363213008835302151 . . . .
If only we could gain a closed form for this special radius, we might be able to guess
the nature of the whole trajectory!
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
19
4. Recent Examples Relating to Our Own Work
(a) Critical Temperature
(b) Wolfram Player Demonstration
Figure 5. The 2-dimensional Ising Model of Ferromagnetism (a) image provided by Jacques Perk, plotting magnetization M (blue) and
specific heat C (red) per site against absolute temperature T [43, p.
91-93, p. 245].
Example 4.1 (Ising integrals [5, 8]). We recently studied the following classes of
integrals [5]. The Dn integrals arise in the Ising model of mathematical physics
(showing ferromagnetic temperature driven phase shifts see Figure 5 and [31]), and
the Cn have tight connections to quantum field theory [8].
Z ∞
Z
4 ∞
1
du1
dun
Cn =
···
···
P
2
n! 0
u1
un
n
0
j=1 (uj + 1/uj )
Q ui −uj 2
Z ∞
Z ∞
i<j ui +uj
4
du1
dun
···
···
Dn =
P
2
n! 0
u1
un
n
0
j=1 (uj + 1/uj )
!2
Z 1
Z 1
Y uk − uj
En = 2
···
dt2 dt3 · · · dtn ,
u
k + uj
0
0
1≤j<k≤n
Qk
where (in the last line) uk = i=1 ti .
20
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
Needless to say, evaluating these multidimensional integrals to high precision
presents a daunting computational challenge. Fortunately, in the first case, the Cn
integrals can be written as one-dimensional integrals:
2n
=
n!
Cn
Z
∞
pK0n (p) dp,
0
where K0 is the modified Bessel function. After computing Cn to 1000-digit accuracy
for various n, we were able to identify the first few instances of Cn in terms of wellknown constants, e.g.,
C3 = L−3 (2) :=
X
n≥0
1
1
−
2
(3n + 1)
(3n + 2)2
,
C4 =
7
ζ(3),
12
where ζ denotes the Riemann zeta function. When we computed Cn for fairly large
n, for instance
C1024 = 0.63047350337438679612204019271087890435458707871273234 . . . ,
we found that these values rather quickly approached a limit. By using the new
edition of the Inverse Symbolic Calculator 7 this numerical value was identified as
lim Cn = 2e−2γ ,
n→∞
where γ is the Euler constant, see Section 5. We later were able to prove this fact—
this is merely the first term of an asymptotic expansion—and thus showed that the
Cn integrals are fundamental in this context [5].
The integrals Dn and En are much more difficult to evaluate, since they are not
reducible to one-dimensional integrals (as far as we can tell), but with certain symmetry transformations and symbolic integration we were able to symbolically reduce
the dimension in each case by one or two.
In the case of D5 and E5 , the resulting 3-D integrands are extremely complicated
(see Figure 6), but we were nonetheless able to numerically evaluate these to at
least 240-digit precision on a highly parallel computer system. This would have been
impossible without the symbolic reduction. We give the integral in extenso to show
the difference between a humanly accessible answer and one a computer finds useful.
7
Available at http://carma.newcastle.edu.au/isc2/.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
Z
E5
1
Z
1
1
Z
=
0
0
2(1 − x)2 (1 − y)2 (1 − xy)2 (1 − z)2 (1 − yz)2 (1 − xyz)2
0
− 4(x + 1)(xy + 1) log(2) y 5 z 3 x7 − y 4 z 2 (4(y + 1)z + 3)x6 − y 3 z y 2 + 1 z 2 + 4(y+
1)z + 5) x5 + y 2 4y(y + 1)z 3 + 3 y 2 + 1 z 2 + 4(y + 1)z − 1 x4 + y z z 2 + 4z
+5) y 2 + 4 z 2 + 1 y + 5z + 4 x3 + −3z 2 − 4z + 1 y 2 − 4zy + 1 x2 − (y(5z + 4)
+4)x − 1)] / (x − 1)3 (xy − 1)3 (xyz − 1)3 + 3(y − 1)2 y 4 (z − 1)2 z 2 (yz
−1)2 x6 + 2y 3 z 3(z − 1)2 z 3 y 5 + z 2 5z 3 + 3z 2 + 3z + 5 y 4 + (z − 1)2 z
5z 2 + 16z + 5 y 3 + 3z 5 + 3z 4 − 22z 3 − 22z 2 + 3z + 3 y 2 + 3 −2z 4 + z 3 + 2
z 2 + z − 2 y + 3z 3 + 5z 2 + 5z + 3 x5 + y 2 7(z − 1)2 z 4 y 6 − 2z 3 z 3 + 15z 2
+15z + 1) y 5 + 2z 2 −21z 4 + 6z 3 + 14z 2 + 6z − 21 y 4 − 2z z 5 − 6z 4 − 27z 3
−27z 2 − 6z + 1 y 3 + 7z 6 − 30z 5 + 28z 4 + 54z 3 + 28z 2 − 30z + 7 y 2 − 2 7z 5
+15z 4 − 6z 3 − 6z 2 + 15z + 7 y + 7z 4 − 2z 3 − 42z 2 − 2z + 7 x4 − 2y z 3 z 3
−9z 2 − 9z + 1 y 6 + z 2 7z 4 − 14z 3 − 18z 2 − 14z + 7 y 5 + z 7z 5 + 14z 4 + 3
z 3 + 3z 2 + 14z + 7 y 4 + z 6 − 14z 5 + 3z 4 + 84z 3 + 3z 2 − 14z + 1 y 3 − 3 3z 5
+6z 4 − z 3 − z 2 + 6z + 3 y 2 − 9z 4 + 14z 3 − 14z 2 + 14z + 9 y + z 3 + 7z 2 + 7z
+1) x3 + z 2 11z 4 + 6z 3 − 66z 2 + 6z + 11 y 6 + 2z 5z 5 + 13z 4 − 2z 3 − 2z 2
+13z + 5) y 5 + 11z 6 + 26z 5 + 44z 4 − 66z 3 + 44z 2 + 26z + 11 y 4 + 6z 5 − 4
z 4 − 66z 3 − 66z 2 − 4z + 6 y 3 − 2 33z 4 + 2z 3 − 22z 2 + 2z + 33 y 2 + 6z 3 + 26
z 2 + 26z + 6 y + 11z 2 + 10z + 11 x2 − 2 z 2 5z 3 + 3z 2 + 3z + 5 y 5 + z 22z 4
+5z 3 − 22z 2 + 5z + 22 y 4 + 5z 5 + 5z 4 − 26z 3 − 26z 2 + 5z + 5 y 3 + 3z 4 −
22z 3 − 26z 2 − 22z + 3 y 2 + 3z 3 + 5z 2 + 5z + 3 y + 5z 2 + 22z + 5 x + 15z 2 + 2z
+2y(z − 1)2 (z + 1) + 2y 3 (z − 1)2 z(z + 1) + y 4 z 2 15z 2 + 2z + 15 + y 2 15z 4
−2z 3 − 90z 2 − 2z + 15 + 15 / (x − 1)2 (y − 1)2 (xy − 1)2 (z − 1)2 (yz − 1)2
(xyz − 1)2 − 4(x + 1)(y + 1)(yz + 1) −z 2 y 4 + 4z(z + 1)y 3 + z 2 + 1 y 2
−4(z + 1)y + 4x y 2 − 1 y 2 z 2 − 1 + x2 z 2 y 4 − 4z(z + 1)y 3 − z 2 + 1 y 2
+4(z + 1)y + 1) − 1) log(x + 1)] / (x − 1)3 x(y − 1)3 (yz − 1)3 − [4(y + 1)(xy
+1)(z + 1) x2 z 2 − 4z − 1 y 4 + 4x(x + 1) z 2 − 1 y 3 − x2 + 1 z 2 − 4z − 1
y 2 − 4(x + 1) z 2 − 1 y + z 2 − 4z − 1 log(xy + 1) / x(y − 1)3 y(xy − 1)3 (z−
1)3 − 4(z + 1)(yz + 1) x3 y 5 z 7 + x2 y 4 (4x(y + 1) + 5)z 6 − xy 3 y 2 +
1) x2 − 4(y + 1)x − 3 z 5 − y 2 4y(y + 1)x3 + 5 y 2 + 1 x2 + 4(y + 1)x + 1 z 4 +
y y 2 x3 − 4y(y + 1)x2 − 3 y 2 + 1 x − 4(y + 1) z 3 + 5x2 y 2 + y 2 + 4x(y + 1)
y + 1) z 2 + ((3x + 4)y + 4)z − 1 log(xyz + 1) / xy(z − 1)3 z(yz − 1)3 (xyz − 1)3
/ (x + 1)2 (y + 1)2 (xy + 1)2 (z + 1)2 (yz + 1)2 (xyz + 1)2 dx dy dz
Figure 6. The reduced multidimensional integral for E5 , which integral has led via extreme-precision numerical quadrature and PSLQ to
the conjectured closed form given in (4.1).
21
22
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
In this way, we produced the following evaluations, all of which except the last we
subsequently were able to prove:
D2
D3
D4
E2
E3
E4
=
=
=
=
=
=
1/3
8 + 4π 2 /3 − 27 L−3 (2)
4π 2 /9 − 1/6 − 7ζ(3)/2
6 − 8 log 2
10 − 2π 2 − 8 log 2 + 32 log2 2
22 − 82ζ(3) − 24 log 2 + 176 log2 2 − 256(log3 2)/3 + 16π 2 log 2 − 22π 2 /3
and
?
E5 = 42 − 1984 Li4 (1/2) + 189π 4 /10 − 74ζ(3) − 1272ζ(3) log 2 + 40π 2 log2 2
−62π 2 /3 + 40(π 2 log 2)/3 + 88 log4 2 + 464 log2 2 − 40 log 2,
(4.1)
where Li denotes the polylogarithm function.
In the case of D2 , D3 and D4 , these are confirmations of known results. We tried
but failed to recognize D5 in terms of similar constants (the 500-digit numerical
value is accessible8 if anyone wishes to try to find a closed form; or in the manner
of the hard sciences to confirm our data values). The conjectured identity shown
here for E5 was confirmed to 240-digit accuracy, which is 180 digits beyond the level
that could reasonably be ascribed to numerical round-off error; thus we are quite
confident in this result even though we do not have a formal proof [5].
Note that every one of the D, E forms above, including the conjectured last one,
is hyperclosed in the sense of our “Sixth approach” section.
Example 4.2 (Weakly coupling oscillators [48, 6]). In an important analysis of coupled Winfree oscillators, Quinn, Rand, and Strogatz [48] developed a certain N oscillator scenario whose bifurcation phase offset φ is implicitly defined, with a
conjectured asymptotic behavior: sin φ ∼ 1 − c1 /N , with experimental estimate
c1 = 0.605443657 . . .. In [6] we were able to derive the exact theoretical value of this
“QRS constant” c1 as the unique zero of the Hurwitz zeta ζ(1/2, z/2) on z ∈ (0, 2).
In so doing were able to prove the conjectured behavior. Moreover, we were able
to sketch the higher-order asymptotic behavior; something that would have been
impossible without discovery of an analytic formula.
Does this deserve to be called a closed form? In our opinion resoundingly ‘yes’,
unless all inverse functions such as that in Bornemann’s (3.8) are to be eschewed.
Such constants are especially interesting in light of even more recent work by Steve
Strogatz and his collaborators on chimera—coupled systems which can self-organize
8Available
at http://crd.lbl.gov/~dhbailey/dhbpapers/ising-data.pdf.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
a
π
week ending
22 AUGUST 2008
PHYSICAL REVIEW LETTERS
PRL 101, 084103 (2008)
b
c
23
1
a
0
1
b
0
1
c
r
θj
0
−π
0
500
oscillator index j
1000
t
f (θ)
FIG. 1 (color online). Snapshot of a chimera state, obtained by
numerical integration of (1) with 0:1, A 0:2, and N1 N2 1024. (a) Synchronized population. (b) Desynchronized
population. (c) Density of desynchronized phases predicted by
Eqs. (6) and (12) (smooth curve) agrees with observed histogram.
remains constant, except for slight fluctuations due to
finite-size effects. Thus, this chimera is stable and statistically stationary. However, if we increase (the coupling
within a population) relative to (the coupling between
populations), the stationary state can lose stability. Now the
order parameter pulsates, and the chimera starts to breathe
[Fig. 2(b)]. The breathing cycle lengthens as we increase
the disparity A between the couplings [Fig. 2(c)].
At a critical disparity, the breathing period becomes infinite. Beyond that, the chimera disappears and the synchronized state becomes a global attractor.
To explain these results, we analyze Eq. (1) in the
continuum limit where N ! 1 for 1, 2. Then
Eq. (1) gives rise to the continuity equations
FIG. 2 (color online). Order parameter r versus time. In all
three panels, N1 N2 128 and 0:1. (a) A 0:2: stable
chimera; (b) A 0:28: breathing chimera; (c) A 0:35: longperiod breather. Numerical integration began from an initial
condition close to the chimera state, and plots shown begin after
allowing a transient time of 2000 units.
1
z ei ei z ei ei ;
(5)
2i where the denotes complex conjugate.
Following Ott and Antonsen [11], we now consider a
special class of density functions f that have the form of a
Poisson kernel. The remarkable fact that Ott and Antonsen
discovered is that such kernels satisfy the governing equations exactly, if a certain low-dimensional system of ordinary differential equations is satisfied. In other words, for
this family of densities, the dynamics reduce from infinite
dimensional to finite (and low) dimensional. (Numerical
evidence suggests that all attractors lie in this family, but
proving this remains an open problem.) Specifically, let
X
1
1
1
f ; t a tei n c:c: : (6)
2
n1
v ; t ! Figure 7. Simulated chimera (figures and parameters from [42])
in parts of their domain and remain disorganized elsewhere, see Figure 7 taken from
[42]. In this case observed numerical limits still need to be put in closed form.
It is a frequent experience of ours that, as in Example 4.2, the need for high
accuracy computation drives the development of effective analytic expressions (closed
forms?) which in turn
shed substantial
light on the subject being studied.
@
@f typically
(2)
f v 0;
@
@t
where f ; t is the probability density of oscillators in
population , and v ; t is their velocity, given by
What is special here is that we use the same function a t
the Fourier
except
that a is raised
the
Example 4.3 (Box integrals [3, 7, 22]). Thereinnthallhas
beenharmonics,
recent
research
ontocalculation
power in the nth harmonic. Inserting this f into the
governing equations,
one finds
that this is an exactSuch
solution, expectaof expected distance
of Zpoints inside a hypercube
to the
hypercube.
2
X
as long as
v ; t ! K sin0 f 0 ; td0 :
tions are also called
“box integrals” [22]. So fora_ example,
expectation
h|~r|i for
1 2 the
i
1
z ei 0:
(7)
i!a 2a z e
random ~r ∈ [0, 1]3 has the closed form (3) Instead of infinitely many amplitude equations, we have
(Note that we dropped the superscripts on to ease the
just one. (It is the same equation for all n.)
we express the complex order
notation. Thus, means and 10 means
√ .) If1we define 1 To close the√system,
a complex order parameter
parameter
in terms
3 − π + log
2 z+
3 of. a. Inserting the Poisson kernel
2
24
2 (6) into Eq. (4), and performing the integrations, yields
Z 4
X
i 0
0
0
0
0
0
z t K0
0
0
e f ; td ;
(4)
2
X
z t Kcuriosity—the
(8) integrals
Incidentally,
box integrals are not just a mathematician’s
a t;
1
then v simplifies to
have been used recently to assess the randomness
of
brain
synapses
positioned
within
084103-2
a parallelepiped [37]. Indeed, we had cognate results for
Z
Z
∆d (s) :=
kx − yks2 dxdy
0 1
0
0
0
[0,1]d
[0,1]d
which gives the moments of the distance between two points in the hypercube.
24
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
In a lovely recent paper [51] Stephan Steinerberger has shown that in the limit as
the dimension goes to infinity
s/p
s/p Z
Z
1
2
s
(4.2)
lim
kx − ykp dxdy =
d→∞
d
(p + 1)(p + 2)
[0,1]d [0,1]d
for any s, p > 0. In particular, with p = 2 this gives a first-order answer to our earlier
published request for the asymptotic behavior of ∆d (s).
A quite recent result is that all box integrals h|~r|n i for integer n, and dimensions
1, 2, 3, 4, 5 are hyperclosed, in the sense of our “Sixth attempt” section. It turns
out that five-dimensional box integrals have been especially difficult, depending on
knowledge of a hyperclosed form for a single definite integral J(3), where
Z
log(t + x2 + y 2 )
J(t) :=
dx dy.
(4.3)
2
2
[0,1]2 (1 + x )(1 + y )
A proof of hyperclosure of J(t) for algebraic t ≥ 0 is established in [22, Thm. 5.1].
Thus h|~r|−2 i for ~r ∈ [0, 1]5 can be written in explicit hyperclosed form involving a
105 -character symbolic J(3); the authors of [22] were able to reduce the 5-dimensional
box integral down to “only” 104 characters. A companion integral J(2) also starts out
with about 105 characters but reduces stunningly to a only a few dozen characters,
namely
π 29
π2
7
11
5π
J(2) =
log 2 − ζ(3) +
π Cl2
−
π Cl2
,
(4.4)
8
48
24
6
24
6
P
where Cl2 is the Clausen function Cl2 (θ) := n≥1 sin(nθ)/n2 (Cl2 is the simplest
non-elementary Fourier series).
Automating such reductions will require a sophisticated simplification scheme with
a very large and extensible knowledge base. With a current Research Assistant, Alex
Kaiser at Berkeley, we have started to design software to refine and automate this
process and to run it before submission of any equation-rich paper. This semiautomated integrity checking becomes pressing when—as above—verifiable output
from a symbolic manipulation can be the length of a Salinger novella.
5. Profound curiosities
In our treatment of numbers enjoying hyperclosure or superclosure, we admitted
that such numbers are countable, and so almost all complex numbers cannot be
given a closed form along such lines. What is stultifying is: How do we identify an
explicit number lying outside of such countable sets? The situation is tantamount to
the modern bind in regard to normal numbers—numbers which to some base have
statistically random digit-structure in a certain technical sense. The bind is: Though
almost all numbers are absolutely normal (i.e. normal to every base 2, 3, . . . ), we do
Media • Resources
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
25
edited by
not know a single fundamental constant
that is provably absolutely normal. (We do
Jon L. Holmes
Nancy S. Gettys
know some “artificial” normal numbers,
see [13].)
University of Wisconsin–Madison
Here is one possible way out Madison,
of theWI dilemma:
In the theory of computability, the
53706
existence of noncomputable real numbers, such as an encoded list of halting Turing
uded in Advanced
Chemistry Collection,
machines, is well established. The celebrated Chaitin constant Ω is a well-known
e 28, a CD-ROM
for StudentsSo a “folk” argument goes: “Since every element of the ring of
noncomputable.
hyperclosure H can be computed via converging series, it should be that Ω 6∈ H.” A
good• research
problem
would
be toofmake
this heuristic rigorous.
Animations that
show fractional
occupation
ions
within
a
unit
cell.
Let us focus on some constants that might not be hyperclosed (nor
Pn superclosed).
Plots of lattice energies and electrostatic interactions,
d Justin T. Fermann,
One• such
constant
is
the
celebrated
Euler
constant
γ
:=
lim
n→∞
k=1 1/k − log n.
as well as their relationship to charge, bond length,
sachusetts, Amherst,
We know
no hypergeometric
form
for γ; said constant may well lie outside of H
Born of
exponent,
etc. (Data have been taken
from sevs.edu
eral S).
sources
[1–4].) are expansions for the Euler constant, such as
(or even
There
Lattice Energetics can be used to present theconcepts
Windows computers
X
log(2k + 1)
3
of lattice energetics to a class. It can also be made available 4
ogether, are designed
(−1)k+1
+
,
γ
=
log
π
−
4
log
Γ
for
students
to
explore
on
an
individual
basis
to
help
them
nteractions between
4
π k≥1
2k + 1
learn about lattice energetics and the Madelung constant.
t electrostatic inbereft of the con-
epth look at the
systems such as
The terms of the
res of what they
ical dynamics of
lso presented.
ons. It combines
ergy calculations
tion to teach stuted.
and
even more exotic series (see [12]). But in the spirit of the present treatment, we
Literature Cited
do not want to call the infinite series closed because it is not hypergeometric per se.
1. CRC Handbook of Chemistry and Physics, 63rd edition; CRC
Relatedly,
the classical Bessel expansion is
Press: Boca Raton, Florida, 1982; pp D-52–D-95.
∞ Pn−1 1 2 n
z Reference Da2. NIST Chemistry WebBook, NIST
(Standard
X
z
k=1 k
tabase NumberK
69–November
1998 Release),
+ γhttp://
I0 (z) +
.
0 (z) = − ln
2
webbook.nist.gov/chemistry/ (accessed Nov 2002).
2
(n!)
4
n=1
3. Masterton, W. L.; Hurley, C. N. Chemistry: Principles and Re-
actions;
Saunders College Publishing: New York, 1989.
Now K
0 (z) has a (degenerate) Meijer-G representation—so potentially is superclosed
4. West, A. R. Solid State Chemistry and its Applications; John
for algebraic
z—and
is accordingly hyperclosed, but the nested-harmonic series
0 (z)1984.
Wiley and Sons
Ltd.: NewIYork,
on the right is again problematic. Again, γ is conjectured not to be a period [41].
t was covered in
rences.
ntal concepts about
concepts about the
ystal structures. Fea-
delung series and
ning in the struc-
e relationship bend the calculated
and charge of the
A screen from the Madelung Constant section of Lattice Energetics.
(a) NaCl nearest neighbors
(b) CMS Prize Sculpture
Figure 8. Two representations of salt
• Vol. 80 No. 1 January 2003 • JChemEd.chem.wisc.edu
26
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
Example 5.1 (Madelung constant [21, 36, 57]). Another fascinating number is the
Madelung constant, M, of chemistry and physics [21, Section 9.3]. This is the potential energy at the origin of an oscillating-charge crystal structure (most often said
crystal is NaCl (salt) as illustrated in Figure 8). The right-hand image is of a Helaman Ferguson sculpture based on M that is awarded biannually by the Canadian
Mathematical Society as part of the David Borwein Career Award.) and is given by
the formal (conditionally convergent [17]) sum
X
(−1)x+y+z
p
M :=
= −1.747564594633...,
(5.1)
2 + y2 + z2
x
(x,y,z)6=(0,0,0)
and has never been given what a reasonable observer would call a closed form. Nature
plays an interesting trick here: There are other crystal structures that are tractable,
yet somehow this exquisitely symmetrical salt structure remains elusive. In general,
even dimensional crystal sums are more tractable than odd for the same modular
function reasons that the number of representations of a number as the sum of an
even number of squares is. But this does not make them easy as illustrated by
Example 1.2.
Here we have another example of a constant having no known closed form, yet
rapidly calculable. A classical rapid expansion for the Madelung constant is due to
Benson:
π p
X
M = −12π
sech2
(2m + 1)2 + (2n + 1)2 ,
(5.2)
2
m,n≥0
in which convergence is exponential. Summing for m, n ≤ 3 produces −1.747564594 . . .,
correct to 8 digits. There are great many other such formulae for M (see [21, 34]).
Through the analytic methods of Buhler, Crandall, Tyagi and Zucker since 1999
(see [34, 36, 54, 57]), we now know approximations such as
1 log 2 8π Γ(1/8)Γ(3/8)
k42
√
−
+
+
+ log
,
8
4π
3
16k4 k40
π 3/2 2
where k4 := ((21/4 − 1)/(21/4 + 1))2 . Two remarkable things: First, this approximation is good to the same 13 decimals we give in the display (5.1); the missing
O(10−14 ) error here is a rapidly, exponentially converging—but alas infinite—sum
in this modern approximation theory. Second: this 5-term approximation itself is
indeed hyperclosed, the only problematic term being the Γ-function part, but we did
establish in our “Sixth approach” section that B(1/8, 3/8) and also 1/π are hyperclosed, which is enough. Moreover, the work of Borwein and Zucker [26] also settles
hyperclosure for that term.
M ≈
Certainly we have nothing like a proof, or even the beginnings of one, that M (or
γ) lies outside H (or even S), but we ask on an intuitive basis: Is a constant such
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
27
as the mighty M telling us that it is not hyperclosed, in that our toil only seems to
bring more “closed-form” terms into play, with no exact resolution in sight?
6. Concluding Remarks and Open Problems
• We have posited several approaches to the elusive notion of “closed form.”
But what are the intersections and interrelations of said approaches? For
example, can our “Fourth approach” be precisely absorbed into the evidently
more general “Sixth approach” (hyperclosure and superclosure)?
• How do we find a single number that is provably not in the ring of hyperclosure
H? (Though no such number is yet known, almost all numbers are as noted
not in said ring!) The same question persists for the ring of hyperclosure, S.
Furthermore, how precisely can one create a field out of HH via appropriate
operator extension?
• Though H is a subset of S, how might one prove that H 6= S? (Is the inequality
even true?) Likewise, is the set of closed forms in the sense of [47, Ch. 8] (only
finite linear combinations of hypergeometric evaluations) properly contained
H
in our H? And what about a construct such as HH ? Should such an entity
be anything really new? Lest one remark on the folly of such constructions,
π
we observe that most everyone would say π π is a closed form!
• Having established the property of hyperclosure for Γb (a/b), are there any
cases where the power b may
√ be brought down? For example, 1/π is hyperclosed, but what about 1/ π?
• What is a precise connection between the ring of hyperclosure (or superclosure) and the set of periods or of Mahler measures (as in Example 1.3)?
• There is expounded in reference [22] a theory of “expression entropy,” whereby
some fundamental entropy estimate gives the true complexity of an expression.
So for example, an expression having 1000 instances of the polylog token
Li3 might really involve only about 1000 characters, with that polylogarithm
token encoded as a single character, say. (In fact, during the research for
[22] it was noted that the entropy of Maple and Mathematica expressions of
the same entity often had widely varying text-character counts, but similar
entropy assessments.)
On the other hand, one basic notion of “closed form” is that explicitly infinite sums not be allowed. Can these two concepts be reconciled? Meaning:
Can we develop a theory of expression entropy by which an explicit, infinite
sum
infinite entropy? This might be difficult, as for example a sum
P∞ is given
1
n=1 n3/2 only takes a few characters to symbolize, as we just did hereby! If
28
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
one can succeed, though, in resolving thus the entropy business for expressions, “closed form” might be rephrased as “finite entropy.”
In any event, we feel strongly that the value of closed forms increases as the
complexity of the objects we manipulate computationally and inspect mathematically
grows—and we hope we have illustrated this. Moreover, we belong to the subset of
mathematicians which finds fun in finding unanticipated closed forms.
Acknowledgements. Thanks are due to David Bailey and Richard Brent for many
relevant conversations and to Armin Straub for the complex plots of W3 and W4 .
References
[1] G. E. Andrews and R. Askey and R. Roy (1999). Special Functions, Cambridge University
Press.
[2] D.H. Bailey, R. Barrio, and J.M. Borwein (2010). “High performance computation:
mathematical physics and dynamics.” Prepared for Joint SIAM/RSME-SCM-SEMA Meeting
on Emerging Topics in Dynamical Systems and Partial Differential Equations (DSPDEs’10)
May 31–June 4, 2010. Available at http://carma.newcastle.edu.au/jon/hpmd.pdf.
[3] D. Bailey, J. Borwein, N. Calkin, R. Girgensohn, R. Luke and V. Moll (2007). Experimental
Mathematics in Action, A K Peters, Natick, MA.
[4] D. H. Bailey and J. M. Borwein (2008). “Computer-assisted discovery and proof.” Tapas in
Experimental Mathematics, 21–52, in Contemporary Mathematics, 457, American
Mathematical Society, Providence, RI, 2008.
[5] D. H. Bailey, D. Borwein, J.M. Borwein and R.E. Crandall (2007). “Hypergeometric forms
for Ising-class integrals.” Experimental Mathematics, 16, 257–276.
[6] D.H. Bailey, J.M, Borwein and R.E. Crandall (2008) “Resolution of the
Quinn–Rand–Strogatz constant of nonlinear physics.” Experimental Mathematics, 18,
107–116.
[7] D. H. Bailey, J. M. Borwein and R. E. Crandall (2010). “Advances in the theory of box
integral.” Mathematics of Computation, E-published February 2010.
[8] D.H. Bailey, J.M. Borwein, D.M. Broadhurst and L. Glasser (2008). “Elliptic integral
representation of Bessel moments.” J. Phys. A: Math. Theory, 41 5203–5231.
[9] D. H. Bailey, J. M. Borwein, V. Kapoor and E. Weisstein (2006). “Ten Problems in
Experimental Mathematics.” American Mathematical Monthly, 113 481–409.
[10] D. H. Bailey, P. B. Borwein and S. Plouffe (1997). “On the Rapid Computation of Various
Polylogarithmic Constants.” Mathematics of Computation, 66 903–913.
[11] R. Baillie, D. Borwein, and J. Borwein (2008). “Some sinc sums and integrals.” American
Math. Monthly, 115 888–901.
[12] D. Bailey, and R. Crandall (2001). “On the random character of fundamental constant
expansions.” ExperimentaL Mathematics, 10 175–190.
[13] D. Bailey, and R. Crandall (2002).“Random generators and normal number.,” Experimental
Mathematics, 11 527–547.
[14] Folkmar Bornemann (2005). “How Mathematica and Maple get Meijer’s G-function into
Problem 9.” Preprint. Available at
http://www-m3.ma.tum.de/bornemann/Numerikstreifzug/Chapter9/MeijerG.pdf.
CLOSED FORMS: WHAT THEY ARE AND WHY WE CARE
29
[15] F. Bornemann, D. Laurie, S. Wagon, and J. Waldvogel (2004). The Siam 100-Digit
Challenge: A Study In High-accuracy Numerical Computing, SIAM, Philadelphia.
[16] E. J. Borowski and J. M. Borwein (2006). Web-linked Dictionary of Mathematics,
Smithsonian/Collins Edition.
[17] D. Borwein, J.M. Borwein and K.F. Taylor (1985), “Convergence of lattice sums and
Madelung’s constant,” J. Math. Phys., 26, 2999–3009.
[18] J. M. Borwein (2005). “The SIAM 100 Digits Challenge.” Extended review in the
Mathematical Intelligencer, 27 40–48.
[19] J. M. Borwein and D. H. Bailey (2008). Mathematics by Experiment: Plausible Reasoning in
the 21st Century, extended second edition, A K Peters, Natick, MA.
[20] J. M. Borwein, D. H. Bailey and R. Girgensohn (2004). Experimentation in Mathematics:
Computational Roads to Discovery, A K Peters, Natick, MA.
[21] J. M. Borwein and P. B. Borwein (1987). Pi and the AGM, John Wiley.
[22] J. M. Borwein, O-Yeat Chan and R. E. Crandall (2010). “Higher-dimensional box integrals.”
Submitted Experimental Mathematics.
[23] (2010). J. Borwein, D. Nuyens, A. Straub, and James Wan, “Random walk integrals.”
Preprint available at http://carma.newcastle.edu.au/jon/walks.pdf.
[24] J. M. Borwein and K. Devlin (2008). The Computer as Crucible, A K Peters, Natick, MA.
[25] M. Bronstein, R. M. Corless, J.H. Davenport, and D. J. Jeffrey (2008). “Algebraic properties
of the Lambert W function from a result of Rosenlicht and of Liouville.” Integral Transforms
and Special Functions, 19 709–712.
[26] J. M. Borwein and I. J. Zucker (1991). “Fast Evaluation of the Gamma Function for Small
Rational Fractions Using Complete Elliptic Integrals of the First Kind,” IMA J. Numerical
Analysis 12 519–526.
[27] J. Carette (2010). Private communication.
[28] J. Carette (2004). “Understanding expression simplification.” Proceedings of International
Conference on Symbolic and Algebraic Computation, ACM 72–79.
[29] S. Chandrasekhar (1967). An Introduction to the Study of Stellar Structures, Dover, New
York.
[30] T. Y. Chow, (1999). “What is a closed-Form Number?” American Mathematical Monthly,
106 440–448.
[31] B. A. Cipra (1987), “An introduction to the Ising model,” Amer. Math. Monthly, 94 (10),
937–959.
[32] R. E. Crandall (1996). Topics in Advanced Scientific Computation, Springer, New York.
[33] R. E. Crandall (2007). “Theory of ROOF Walks.” Available at
http://people.reed.edu/~crandall/papers/ROOF.pdf.
[34] R. E. Crandall (1999). “New representations for the Madelung constant.” Experimental
Mathematics, 8 367–379.
[35] R.E. Crandall (2009). “Analytic representations for circle-jump moments.” PSIpress, 21 Nov
09, www.perfscipress.com/papers/analyticWn_psipress.pdf
[36] R. E. Crandall and J. E. Buhler (1987). “Elementary expansions for Madelung constants.” J.
Phys. A: Math Gen. 20 5497–5510.
[37] R. Crandall and T. Mehoke (2010). “On the fractal distribution of brain synapses.” Preprint.
[38] C. Deninger (1997) “Deligne periods of mixed motives, theory and the entropy of certain
actions.” J. Amer. Math. Soc., 10 259–281.
30
JONATHAN M. BORWEIN AND RICHARD E. CRANDALL
[39] H. R. P. Ferguson, D. H. Bailey and S. Arno (1999). “Analysis of PSLQ, An Integer Relation
Finding Algorithm.” Mathematics of Computation, vol. 68, no. 225 (Jan 1999), pg. 351–369.
[40] Brian Hayes (2005). “Why W?” American Scientist, 93 1004–1008.
[41] M. Kontsevich and D. Zagier (2001). “Periods.” In Mathematics unlimited—2001 and beyond,
771–808, Springer-Verlag.
[42] E. A. Martens, C. R. Laing and S. H. Strogatz (2010). “Solvable model of spiral wave
chimeras.” Physical Review Letters, 104 044101.
[43] B. McCoy and Tai Tsun Wu (1973), The Two-Dimensional Ising Model, Harvard Univ Press.
[44] F. W. J. Olver, D. W. Lozier, R. F. Boisvert, and C. W. Clark (2010). NIST Handbook of
Mathematical Functions, Cambridge University Press.
[45] B. D. Nener, N. Fowkes, and L. Borredon (2003). “Analytical models of optical refraction in
the troposphere.” JOSA, 20 867–875.
[46] K. Oldham, J. Myland, and J. Spanier (2009). An Atlas of Functions: With Equator, The
Atlas Function Calculator, Springer-Verlag.
[47] M. Petkovsek, H. Wilf and D. Zeilberger (1996). A K Peters, Ltd. Available at
http://www.math.upenn.edu/~wilf/AeqB.html.
[48] D. Quinn, R. Rand, and S. Strogatz (2007). “Singular unlocking transition in the Winfree
model of coupled oscillators.” Phys. Rev E, 75 036218-1-10.
[49] M. Rogers (2010). “Hypergeometric formulas for lattice sums and Mahler measures.”
Preprint.
[50] C. J. Smyth (2003). “Explicit Formulas for the Mahler Measures of Families of Multivariable
Polynomials.” Preprint.
[51] S. Steinerberger (2010). “Extremal uniformly distributed sequences and random chord
lengths.” Acta Mathematica Hungarica. In press.
[52] C. J. Smyth (2008). “The Mahler measure of algebraic numbers: a survey.” Conference
Proceedings, University of Bristol, 3-7 April 2006. LMS Lecture Note Series 352, Cambridge
University Press, Cambridge, pp. 322–349.
[53] N. Temme (1996). Special Functions, an Introduction to the Classical Functions of
Mathematical Physics, John Wiley & Sons, New York.
[54] S. Tyagi (2003). Private communication.
[55] E. W. Weisstein, “Closed-Form Solution.” From MathWorld – A Wolfram Web Resource.
http://mathworld.wolfram.com/Closed-FormSolution.html.
[56] B. L. Walden and L. A. Ward (2007). “A harmonic measure interpetation of the
arithmetic-geometric mean.” American Mathematical Monthly, 114 610–622.
[57] I. J. Zucker (1984). “Some infinite series of exponential and hyperbolic functions.” SIAM J.
Math. Anal., 15 406–413.